Hi, I'm Sumit

Tokyo-based SWE, aspiring ML Researcher

Preface

Thanks for checking in! I go into some ramblings and random tangents about myself here, so if you want to read a little bit about me, that's a good place.

Right now, I'm particularly interested in LLMs (mostly architectures + finetuning), MlSys, and low-resource language research[1].

Lately...

These days, I've been primarily busy with:

Low resource language research: building datasets, effective benchmarking, tokenization methods, and grants to train models to understand low resource languages of my country. For details, see here.
Learning MLSys: kernel optimizations, HPC, inference serving, etc.
Lots of weekly hackathons and ML events in Tokyo
and doing multiple projects with relevance to the above, publishing blog posts as well as worklogs :)

Looking for Mentorship

I'm currently a solo researcher, and I believe that mentorship in ML research could be mutually valuable! I'm particularly interested in LLM training, inference optimization, fine-tuning, and low-resource language research, including but not limited to dataset curation, tokenization strategies, cross-lingual transfer, and model adaptation for underrepresented languages. If you think some of my research interests or efforts are worthwhile for mentorship or collaboration, please reach out!

Now with that being said, I'm genuinely open to exploring other areas of deep learning if the problems are compelling and the learning opportunity is strong (which might be pretty much every time if you're doing something interesting). I'm still figuring out what I don't know right now, which is exactly why I want to work alongside someone doing real research.

My battle-tested method to learn has always been to put myself under immense pressure from where the only option to survive is via quick learning and adaptation. I love learning by doing; I love working on real problems, asking questions, being wrong, and building up intuition alongside someone who knows what they're doing. Therefore, if you're working on interesting problems and think having someone with relentless curiosity and motivation alongside you could be mutually valuable, I'd love to chat. All I ask for is an environment where I can garner knowledge, skills, and display my grit & toil. As long as I can go all-out in my pursuit of uncompromising learning sense, I don't care about any other gains or rewards.

Email or Twitter work best for me; I'll get back to you as soon as I can.

Hmm... what else?

You can also check out my blog or research pages for project updates and worklogs, or just throw me a message on the aforementioned links if you wanna chat about stuff :)

[1]Most NLP research focuses on English and a handful of high-resource languages, but when I look at the AI status quo in my own country, I see a gap that needs to be filled, and as somebody privileged enough to have access to more education and resources than many, I feel like it's only right that I do my part. I'm particularly fascinated by how we can leverage recent AI breakthroughs to build better technologies for low-resource languages (like the languages of my own country) without massive training data. Time will tell, but I believe lots could be achieved through better cross-lingual transfer, smarter approaches to limited data, and systems that respect linguistic diversity. I do not believe creating synthetic data (like we usually do for high-resource languages) is a way, however, as this will most likely cause translationese effects.